Space-air-ground-forest fire monitoring method and system based on bayesian variational inference

Abstract

The application provides a space-air-ground-forest fire situation monitoring method and system based on Bayesian variational inference, belonging to the technical field of forest fire situation monitoring. The method comprises the following steps: firstly, performing space-time registration on multi-source heterogeneous data, generating a multi-modal feature vector for a grid node, and organizing a dynamic space-time graph. An output and its uncertainty of a physical fire behavior model are used to construct a prior distribution. Bayesian variational inference is performed based on a factor graph and an expectation propagation graph network to obtain a fire situation posterior estimate, an uncertainty measure, and a model cognitive bias. A cognitive conflict degree is obtained by calculating the node-by-node KL divergence between the prior and the posterior, which is used to adjust the graph structure and re-infer, thereby updating the uncertainty and the bias. A joint risk field is further constructed to automatically identify a mutation blind area and trigger a mobile reconnaissance platform to actively supplement data. The supplemented data is re-injected into the graph node to form a closed loop until the blind area is eliminated, and adaptive and accurate monitoring of the fire situation is achieved.

Description

Technical Field

[0001] This application relates to the field of forest fire monitoring technology, and in particular to a method and system for monitoring forest fires in space, air, and ground based on Bayesian variational inference. Background Technology

[0002] Current forest fire monitoring employs a multi-source collaborative architecture involving air, space, and ground. Existing technical solutions process satellite thermal infrared threshold determination, UAV image recognition, and fixed threshold alarms from ground sensors independently, displaying the results in a unified overlay. However, conflicting conclusions from multiple sources rely on manual resolution. Some improved solutions introduce empirical fire spread formulas, combining static parameters such as terrain, weather, and combustible materials to predict fire intensity, and scheduling UAVs for blind spot inspections according to fixed grids and periods. The core flaw lies in its complete reliance on deterministic thresholds and static rules. All parameters and monitoring strategies are pre-fixed during deployment, lacking real-time perception and adaptive adjustment capabilities for the dynamic evolution of the fire. Under complex conditions such as rapid fire expansion or sudden weather changes, the cognitive discrepancy between static predictions and actual fire conditions accumulates continuously. Fixed inspection mechanisms cannot dynamically respond to monitoring blind spots, causing the accuracy and reliability of fire situation assessment to decline over time. This makes it difficult to meet the stringent requirements of modern forest area refined and intelligent prevention and control for real-time situational awareness and proactive risk intervention. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides a method and system for monitoring forest fires in space, air, and ground based on Bayesian variational inference.

[0004] A first aspect of this application provides a method for monitoring forest fires in air, space, and ground based on Bayesian variational inference, including: Spatial registration and temporal alignment are performed on the multi-source heterogeneous data of the target forest area to generate multimodal feature vectors for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data and time series data of ground IoT sensors; A dynamic spatiotemporal graph is constructed using grid nodes. A prior probability distribution is generated based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model. The graph structure of the dynamic spatiotemporal graph is determined based on the prior probability distribution. The dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node are input into a preset graph network based on factor graph and expectation propagation, and the posterior probability distribution is obtained by Bayesian variational inference. Based on the posterior probability distribution, the fire status estimate, uncertainty measure, and model cognitive bias of the physical fire behavior model for each grid node are obtained. The cognitive conflict degree is determined by the node-wise KL divergence between the posterior and prior probability distributions, and the graph structure of the dynamic spatiotemporal graph is adjusted accordingly. Based on the dynamic spatiotemporal graph with the adjusted graph structure, Bayesian variational inference is re-executed to obtain updated uncertainty measures and updated model cognitive biases; A joint risk field is constructed based on the updated uncertainty measure and the updated model cognitive bias, and mutation blind spots that meet the preset mutation conditions are identified in the joint risk field. Based on the mutation blind zone, an active supplementary sampling command is generated to control the mobile reconnaissance platform to collect supplementary data. The supplementary data is updated to the corresponding grid nodes of the dynamic spatiotemporal map, and the complete closed-loop process, from adjusting the graph structure of the dynamic spatiotemporal map according to the degree of cognitive conflict to identifying mutation blind spots, is repeated until the mutation blind spots are resolved.

[0005] A second aspect of this application provides a space-air-ground forest fire monitoring system based on Bayesian variational inference, comprising: The data preprocessing module is used to perform spatial registration and temporal alignment on multi-source heterogeneous data of the target forest area, and generate multimodal feature vectors for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data and time series data of ground IoT sensors; The prior graph construction module is used to construct a dynamic spatiotemporal graph with grid nodes. Based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model, it generates the prior probability distribution of the fire status of each grid node and determines the graph structure of the dynamic spatiotemporal graph based on the prior probability distribution. The Bayesian inference engine is used to input the dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node into a preset graph network based on factor graph and expectation propagation, and to calculate the posterior probability distribution representing the fire status of each grid node through Bayesian variational inference. The cognitive bias assessment module is used to obtain the fire status estimate, uncertainty measure, and model cognitive bias of each grid node based on the posterior probability distribution. The graph structure adaptive module is used to use the node-by-node KL divergence between the posterior probability distribution and the prior probability distribution as the cognitive conflict degree, and to adjust the graph structure of the dynamic spatiotemporal graph according to the cognitive conflict degree. The Bayesian inference engine is also used to re-execute Bayesian variational inference based on the dynamic spatiotemporal graph after adjusting the graph structure, to obtain an updated uncertainty measure and an updated model cognitive bias. The risk field construction and identification module is used to construct a joint risk field based on the updated uncertainty metric and the updated model cognitive bias, and to identify mutation blind spots that meet the preset mutation conditions in the joint risk field. The proactive supplementary sampling decision module is used to generate proactive supplementary sampling instructions based on the sudden change blind zone, and control the mobile reconnaissance platform to collect supplementary sampling data; The closed-loop iterative control module is used to update the grid nodes corresponding to the dynamic spatiotemporal graph with the supplementary data, and trigger the graph structure adaptation module, Bayesian inference engine, risk field construction and identification module and active supplementary data acquisition decision module to repeatedly execute the complete closed-loop process from adjusting the graph structure of the dynamic spatiotemporal graph according to the degree of cognitive conflict to the identification of mutation blind zone, until the mutation blind zone is resolved.

[0006] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described air-space-ground forest fire monitoring method based on Bayesian variational inference.

[0007] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for monitoring forest fires in space, air, and ground based on Bayesian variational inference.

[0008] The beneficial effects of the air-space-ground forest fire monitoring method and system based on Bayesian variational inference provided in this application are as follows: First, this application performs spatial registration and temporal alignment on three types of heterogeneous data: satellite hotspots, UAV videos, and ground-based IoT data, generating multimodal feature vectors for grid nodes, thus constructing a unified observation data base for integrated air-space-ground monitoring. Then, it constructs a priori probability distribution and an initial dynamic spatiotemporal map based on the prediction results of a physical fire behavior model, embedding physical mechanism constraints consistent with the fire spread law into the probabilistic inference. Next, it performs Bayesian variational inference through a factor graph and expectation propagation framework, outputting posterior probability, uncertainty measure, and model cognitive bias to quantify the credibility of monitoring results and the degree of model mismatch. Subsequently, it dynamically adjusts the graph structure using node-by-node KL divergence as the cognitive conflict degree, iteratively optimizing the inference topology. Finally, it identifies abrupt change blind spots through a joint risk field and drives the mobile platform to actively supplement data collection, forming a complete closed loop. This application can adaptively eliminate monitoring blind spots and continuously improve the accuracy and robustness of fire situation estimation. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a method for monitoring forest fires in space, air, and ground based on Bayesian variational inference, provided as an embodiment of this application; Figure 2 A structural block diagram of a space-air-ground forest fire monitoring system based on Bayesian variational inference provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1 -Appendix Figure 3 The following is an explanation using specific examples.

[0012] Please refer to Figure 1 , Figure 1 A flowchart illustrating a method for monitoring forest fires in space, air, and ground based on Bayesian variational inference, provided in an embodiment of this application, is shown below. The method includes: S101: Spatial registration and temporal alignment of multi-source heterogeneous data of the target forest area are performed to generate multimodal feature vectors for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data and time series data of ground IoT sensors.

[0013] In this embodiment, multi-source heterogeneous data is used to comprehensively perceive the fire situation in the target forest area. This multi-source heterogeneous data refers to data from different sources with different formats and sampling frequencies, such as satellite hotspot monitoring data, UAV inspection video stream data, and time-series data from ground-based IoT sensors. Among them, satellite data can provide large-scale thermal anomaly information, UAV video stream data provides high-resolution visual information, and ground sensor data provides local environmental parameters such as temperature and humidity.

[0014] Specifically, data from different sources will have different coordinate systems and sampling frequencies. To achieve data fusion, a geometric transformation algorithm is used to roughly align the spatial positions of the data. Simultaneously, based on the data acquisition timestamps, approximate temporal synchronization is achieved through methods such as linear interpolation or nearest neighbor matching. Subsequently, for each grid node, the original data values ​​or their statistics for all modalities within that node's region are concatenated to form a multimodal feature vector.

[0015] A grid node is an abstract node represented by each grid cell after dividing a target forest area into a series of regular grids. Each grid node carries fire status information and multimodal feature vectors for that area.

[0016] A multimodal feature vector (MMR) is a vector formed by extracting and fusing features from multi-source heterogeneous data that has undergone spatial registration and temporal alignment for each raster node. This MMR represents multiple observational features of the raster node at a specific time step.

[0017] S102: Construct a dynamic spatiotemporal graph using grid nodes. Based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model, generate the prior probability distribution of the fire status of each grid node. Determine the graph structure of the dynamic spatiotemporal graph based on the prior probability distribution. The dynamic spatiotemporal graph contains grid nodes with multiple time steps, and the edges between nodes include spatial adjacency edges and temporal self-connection edges.

[0018] In this embodiment, the dynamic spatiotemporal graph refers to a graph structure constructed with grid nodes as vertices and spatial adjacency and temporal self-connection relationships as edges. The graph structure of this dynamic spatiotemporal graph can be dynamically adjusted to characterize the spatial and temporal evolution of the fire situation, wherein the graph structure of the dynamic spatiotemporal graph contains grid nodes at multiple time steps.

[0019] A physical fire behavior model is a mathematical model that simulates the spread and evolution of fires based on physical parameters such as forest topography, meteorology, and combustible material load. This physical fire behavior model outputs a prediction of fire behavior status and a measure of its uncertainty, providing a basis for the prior probability distribution of fire status.

[0020] The prior probability distribution refers to the initial probability estimate of the fire status of each grid node based on the prediction of the physical fire behavior model and preset uncertainty parameters before the arrival of observation data. This prior probability distribution represents the model's initial understanding of the fire status.

[0021] S103: Input the dynamic spatiotemporal map and the multimodal feature vectors corresponding to each grid node into a preset graph network based on factor graph and expectation propagation, and obtain the posterior probability distribution representing the fire status of each grid node through Bayesian variational inference.

[0022] In this embodiment, Bayesian variational inference refers to an approximate Bayesian inference method that approximates the true posterior probability distribution by optimizing the variational distribution. A factor graph is a bipartite graph that includes variable nodes and factor nodes. Variable nodes represent random variables, and factor nodes represent the functional relationships or probability distributions between these random variables. In graph networks, factor graphs are used to represent the decomposition structure of the joint probability distribution. Expectation propagation refers to an approximate inference algorithm that performs message passing on the factor graph. This algorithm approximates the marginal posterior distribution of the variables by iteratively updating the messages between factors and variables.

[0023] The posterior probability distribution refers to the update probability estimate of the fire status of each grid node, calculated through Bayesian variational inference based on the prior probability distribution and multimodal feature vectors (observation data). This posterior probability distribution represents a more accurate understanding of the fire status.

[0024] S104: Based on the posterior probability distribution, the fire status estimate, uncertainty measure, and model cognitive bias of the physical fire behavior model are obtained for each grid node; the model cognitive bias is the difference between the expectation of the posterior probability distribution and the expectation of the prior probability distribution.

[0025] In this embodiment, model cognitive bias refers to the difference between the expected value of the posterior probability distribution and the expected value of the prior probability distribution. This model cognitive bias quantifies the degree of difference between the predictions of the physical fire behavior model and the actual observed data.

[0026] Specifically, the fire situation estimate is directly taken as the expected value of the posterior probability distribution. The uncertainty measure is taken as the variance of the posterior probability distribution. The model cognitive bias is calculated as the difference between the expected value of the posterior probability distribution and the expected value of the prior probability distribution.

[0027] S105: The cognitive conflict degree is determined by the node-by-node KL divergence between the posterior and prior probability distributions, and the graph structure of the dynamic spatiotemporal graph is adjusted based on the cognitive conflict degree.

[0028] In this embodiment, cognitive conflict degree refers to the KL divergence (i.e., calculated node-by-node) between the posterior probability distribution and the prior probability distribution of each grid node. This cognitive conflict degree characterizes the degree to which the model updates its understanding of the fire status of a node after receiving new observation data, or the degree of conflict with the initial understanding.

[0029] S106: Re-execute Bayesian variational inference based on the dynamic spatiotemporal graph after adjusting the graph structure to obtain the updated uncertainty measure and the updated model cognitive bias.

[0030] In this embodiment, Bayesian variational inference is re-executed based on the dynamic spatiotemporal graph with the adjusted graph structure to obtain an updated uncertainty measure and an updated model cognitive bias. This process follows the steps of Bayesian variational inference described above, running the inference process again on the adjusted graph structure to calculate the updated uncertainty measure and the model cognitive bias recalculated based on the same prior distribution, i.e., the difference between the updated posterior expectation and the original prior expectation.

[0031] It should be noted that in this embodiment, adjusting the graph structure only affects the smoothing factor (i.e., the constraint relationships between nodes) and does not change the prior probability distribution itself. The prior distribution is determined solely by the physical fire behavior model and remains fixed during the iteration process. Therefore, the cognitive bias of the subsequently updated model is still calculated as the difference between the same prior expectation and the updated posterior expectation.

[0032] S107: Construct a joint risk field based on the updated uncertainty measure and the updated model cognitive bias, and identify mutation blind spots that meet the preset mutation conditions in the joint risk field.

[0033] In this embodiment, the joint risk field refers to the spatial distribution of fire risk levels at each grid node in the forest area, constructed based on the updated uncertainty metric and model cognitive bias. This joint risk field comprehensively considers both the uncertainty of the fire situation and the bias in model predictions.

[0034] A mutation blind zone refers to a high-risk connected region identified in the joint risk field that meets preset mutation conditions. This high-risk connected region indicates an area where the fire situation is changing rapidly or where model predictions have significant uncertainties, requiring proactive supplementary sampling.

[0035] Specifically, the joint risk field is constructed as follows: For each grid node i, its updated uncertainty metric Ui and updated model cognitive bias Di are obtained, and the joint risk value Ri is calculated using the following linear weighted formula: Ri = wU·Ui / Umax + wD·(|Di| / Dmax), where |Di| is the absolute value of the model cognitive bias; Umax and Dmax are the maximum values ​​of the absolute values ​​of uncertainty metric and model cognitive bias among all current grid nodes, respectively; wU and wD are preset weight coefficients, for example, wU = 0.4 and wD = 0.6, which are determined by the analytic hierarchy process based on the domain experience that fire monitoring is more sensitive to model mismatch than to noise.

[0036] The inherent relationship in this weighted formula is as follows: the uncertainty measure Ui reflects the credibility of the fire status estimate based on the observation data. The larger the Ui, the greater the conflict between multi-source data or the stronger the observation noise, and the more unreliable the monitoring results in the area. The model cognitive bias Di reflects the systematic difference between the physical fire behavior model prediction and the observation data. The larger the absolute value of Di, the more serious the deviation of the actual fire evolution from the physical model prediction, and the higher the risk of model failure in the area. The weighted sum of these two values ​​allows the joint risk value Ri to comprehensively characterize the unreliability of node i and the degree of model mismatch. The two dimensions of risk complement each other, forming a quantitative expression of a complete risk profile of the fire situation. The Ri values ​​of all grid nodes constitute the joint risk field, represented as a two-dimensional matrix R = {Ri|i∈all grid nodes}.

[0037] S108: Generates active data collection instructions based on the mutation blind zone, and controls the mobile reconnaissance platform to collect supplementary data.

[0038] In this embodiment, the active data acquisition instruction refers to the mission instruction generated for the mobile reconnaissance platform based on the identified abrupt change blind spots. This active data acquisition instruction guides the platform to go to a specific area to collect supplementary data in order to eliminate the blind spots.

[0039] A mobile reconnaissance platform refers to mobile devices such as drones and ground robots used to perform proactive data collection missions. This mobile reconnaissance platform can travel to designated areas according to instructions and collect high-resolution, timely supplementary data.

[0040] For example, firstly, the grids covered by the mutation blind zones in the joint risk field are mapped to access costs based on their joint risk values. Similarly, the areas outside the blind zones are mapped to access costs based on risk values, and these are then merged to generate a composite cost map. The current location, remaining endurance, and current wind field vector of the mobile reconnaissance platform are obtained. For all identified mutation blind zones, the grid node with the highest joint risk value within each blind zone is selected as a candidate target point. The estimated cost of the risky feasible path corresponding to each candidate target point is calculated. When a candidate path exists where the estimated cost is less than or equal to the remaining endurance and the wind field vector is less than or equal to the safe operation threshold, the feasible blind zone with the highest remediation priority index is selected as the mission target. Otherwise (i.e., no candidate path satisfies the above safety and endurance constraints), a multi-attribute evaluation matrix is ​​constructed based on the area, risk exposure value, and remediation urgency of all mutation blind zones. The remediation priority index of each blind zone is calculated using the superior-inferior solution distance method, and feasible mission targets are reassigned from high to low based on this index. The risk exposure value is the sum of the updated joint risk values ​​of all grid nodes within the mutation blind zone. The mobile reconnaissance platform is controlled to perform supplementary data collection along the planned path and the joint risk field is updated in real time within a preset range from the target point. If a new risk intrudes into the planned path, a local replanning is triggered to avoid it.

[0041] S109: Update the supplementary data to the raster nodes corresponding to the dynamic spatiotemporal map, and repeat the complete closed-loop process from adjusting the graph structure of the dynamic spatiotemporal map according to the degree of cognitive conflict to the identification of mutation blind spots, until the mutation blind spots are resolved.

[0042] In this embodiment, after spatial registration and temporal alignment of the supplementary data, the supplementary multimodal feature vector of the raster node is generated according to the same feature extraction and standardization process as in segment S101. Then, it is fused with the original feature vector using an exponentially weighted moving average: Fnew=A·Fs+(1-A)·Fo, where Fnew is the updated raster node multimodal feature vector; Fs is the supplementary multimodal feature vector corresponding to the supplementary data; Fo is the multimodal feature vector; A is the fusion weight coefficient of the supplementary multimodal feature vector, which is dynamically set according to the confidence level of the source of the supplementary data. The higher the confidence level, the larger A is. The initial default A=0.7.

[0043] The criteria for resolving mutation blind zones are that the following three conditions are met simultaneously: (1) the number of mutation blind zones identified in the current iteration is zero; (2) the cognitive conflict degree (node-by-node KL divergence) of all grid nodes is less than 0.5 times the cognitive conflict threshold; (3) the average uncertainty measure of all grid nodes decreases by more than 30% compared with the previous closed-loop iteration, or the absolute value of the change rate of the average uncertainty measure in three consecutive iterations is less than 5%. When the three conditions are met simultaneously, it is considered that the fire situation has become stable and the model cognition has fully converged, and the closed-loop iteration is terminated. If the number of iterations exceeds the preset upper limit (set to 20 times) and the resolution condition is still not met, the system will be forcibly terminated and the current result will be output. At the same time, a system alarm will be issued to prompt manual intervention for judgment.

[0044] As can be seen from the above, this application first performs spatial registration and temporal alignment on three types of heterogeneous data: satellite hotspots, UAV videos, and ground-based IoT data, generating multimodal feature vectors for grid nodes, thus constructing a unified observation data base for integrated air-space-ground monitoring. Next, it constructs a priori probability distribution and an initial dynamic spatiotemporal map based on the prediction results of a physical fire behavior model, embedding physical mechanism constraints consistent with the fire spread law into probabilistic inference. Then, it performs Bayesian variational inference through a factor graph and expectation propagation framework, outputting posterior probability, uncertainty measure, and model cognitive bias to quantify the credibility of monitoring results and the degree of model mismatch. Subsequently, it dynamically adjusts the graph structure using node-by-node KL divergence as the cognitive conflict degree, iteratively optimizing the inference topology. Finally, it identifies abrupt blind spots through a joint risk field and drives the mobile platform to actively supplement data collection, forming a complete closed loop. This application can adaptively eliminate monitoring blind spots and continuously improve the accuracy and robustness of fire situation estimation.

[0045] In one embodiment of this application, a prior probability distribution of the fire status of each grid node is generated based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model. The graph structure of the dynamic spatiotemporal graph is then determined based on the prior probability distribution, including: The digital elevation model, the meteorological forecast wind field, and the spatial distribution of combustible load are input into the physical fire behavior model for simulation. The input parameters are perturbed by Monte Carlo ensemble simulation to obtain the mean and variance of the fire intensity prediction for each grid node. The mean and variance of the fire intensity prediction are normalized, and the prior probability distribution of the fire status of the grid node is constructed using the normalized mean and variance, which is characterized by a beta distribution. For each pair of spatially adjacent grid nodes, calculate the overlap coefficient between the two corresponding prior probability distributions; if and only if the overlap coefficient is greater than the preset overlap threshold, establish a directed spatial adjacency edge between the pair of spatially adjacent grid nodes, and set the edge weight of the directed spatial adjacency edge as the overlap coefficient; wherein, the direction of the directed spatial adjacency edge is from the node with the higher expected value of the prior probability to the node with the lower expected value. For each grid node, the fire intensity prediction sample sequence generated at adjacent time steps is simulated using Monte Carlo set simulation. The Pearson correlation coefficient of the fire intensity prediction sample sequence is calculated, and the absolute value is taken as the time-series autocorrelation coefficient. A time-series self-connecting edge is established between adjacent time steps, and the edge weight is set as the time-series autocorrelation coefficient.

[0046] In this embodiment, the digital elevation model (DEM), forecast wind field, and spatial distribution of combustible load are key parameters describing the environmental characteristics of forest areas, directly influencing the spread of fires. The DEM provides information on topographic relief, affecting the upward or downward propagation speed of the fire; the forecast wind field determines the direction and speed of fire spread; and the spatial distribution of combustible load reflects the quantity and type of fire fuel. These data can be acquired in various ways. For example, the DEM can be obtained through satellite remote sensing, aerial photogrammetry, or lidar scanning; the forecast wind field can be derived from real-time observation data from regional meteorological stations, the output of numerical weather prediction models, or through inversion from meteorological satellite data; and the spatial distribution of combustible load can be obtained through a combination of forestry resource surveys, remote sensing image classification, and biomass estimation models.

[0047] The physical fire behavior model in this embodiment adopts a fire spread simulation architecture based on cellular automata. This architecture defines each grid node as a cell, and the fire behavior state of each cell at time t is represented by a vector consisting of three physical quantities: fire intensity (kW / m), flame length (m), and spread rate (m / min).

[0048] Cell state transitions follow the Rothermel spread equation: R = R0·(1+φw+φs), where R0 is the baseline spread velocity (obtained from a table based on the combustible material model), φw is the wind speed correction coefficient (calculated by accelerating the wind field from the weather forecast and the terrain), and φs is the slope correction coefficient (obtained by calculating the slope aspect and slope angle from the digital elevation model).

[0049] The fire intensity of a cell is calculated using the Byram formula: I = H·w·R, where H is the lower heating value of the combustible (MJ / kg, obtained from the combustible type table by looking up the spatial distribution of combustible load), w is the effective combustible load (t / ha), and R is the spread rate.

[0050] The thermal radiation conduction between cells adopts the Radial radiation model, which states that the radiation flux received by a cell is directly proportional to the fire intensity of its neighboring cells and inversely proportional to the square of the distance. The fire spread between adjacent cells is triggered when the total radiation flux received by the adjacent cells exceeds the critical ignition flux of the combustible material in that cell (preset to 50 kW / m²), and the current temperature of that cell exceeds the ignition temperature (preset to 300°C). The model time step adopts an adaptive step size strategy, with a step size Δτ = min(Δx / vmax, 1.0 s), where Δx is the grid side length and vmax is the maximum spread velocity among all cells.

[0051] Monte Carlo ensemble simulation perturbs the input parameters by randomly sampling and varying the input parameters of the physical fire behavior model within a certain range, thereby generating a series of different simulation scenarios. These input parameters include wind speed, wind direction, and combustible material moisture content. The perturbation used in this embodiment captures the uncertainty of the input parameters, making the fire intensity prediction output by the model more robust and capable of representing multiple possibilities in fire evolution. For example, a uniformly distributed perturbation is applied to the wind speed within ±10% of the predicted value, a random perturbation is applied to the wind direction within ±15 degrees of the predicted direction, and a ±15% normal distribution perturbation is applied to the combustible material moisture content. Through multiple simulations, the predicted fire intensity value for each grid node under different perturbation conditions is obtained, and then the mean and variance of the predicted fire intensity are calculated, representing the average intensity of the fire and its uncertainty measure, respectively.

[0052] The predicted mean and variance of fire intensity are normalized to transform these physical quantities into a uniform interval of [0,1], facilitating the subsequent construction of probability distributions. The normalization method is Z-score normalization (converting the data into a distribution with a mean of 0 and a standard deviation of 1). The prior probability distribution of the fire status of each grid node, characterized by a beta distribution, is constructed using the normalized mean and variance. This is because the beta distribution, defined as a continuous probability distribution in the interval [0,1], is well-suited for representing probabilities or proportions and can effectively characterize the likelihood and uncertainty of fire occurrence. For example, using the moment matching method, the two shape parameters α and β of the beta distribution are calculated based on the normalized mean and variance, thereby determining the prior probability distribution of the fire status for each grid node.

[0053] The specific construction method is as follows: Let the normalized mean be μ, and the variance be σ. 2 The two shape parameters α and β of the beta distribution are solved using the moment matching method. The formula is α = μ·(μ(1-μ) / σ 2 -1), β=(1-μ)·(μ(1-μ) / σ 2 -1). When σ 2When ≥μ(1-μ), i.e. the variance is too large and the parameter becomes invalid, it is forcibly corrected to σ. 2 =μ(1-μ)·0.95, making both α and β positive. For extreme cases where the predicted mean of fire intensity is close to 0 or 1, μ is subjected to boundary shrinkage: if μ<0.01, then set μ=0.01; if μ>0.99, then set μ=0.99, to prevent the shape parameter from being too small and causing numerical instability.

[0054] The overlap coefficient is an indicator that measures the similarity between two probability distributions, and in this context, it is used to quantify the correlation between the fire states of adjacent grid nodes. A higher overlap between the prior fire probability distributions of two grid nodes indicates a greater likelihood of fire propagation between them. The overlap coefficient is calculated using the Batachalia coefficient, or by calculating the proportion of the intersection area of ​​the two distributions to the total area. A preset overlap threshold is the standard for determining whether to establish a spatial adjacency edge between adjacent grid nodes. This preset overlap threshold is determined as follows: fire line intensity data of adjacent grid nodes in historical fire spread events in the target forest area are collected; the overlap coefficients of the fire states of adjacent nodes at different time steps are calculated; and the correlation between these overlap coefficients and the actual fire spread direction is statistically analyzed. Using ROC curve analysis, the true positive rate is defined as the correct connection rate (the proportion of actual fire spread direction consistent with the directed edge direction), and the false positive rate is defined as the incorrect connection rate. The overlap coefficient value corresponding to the maximum Youden exponent is selected as the preset overlap threshold. If historical data is unavailable, relevant fire behavior literature recommends a value based on the main combustible material types in the forest area, such as coniferous forests, broad-leaved forests, and grasslands. This value is initially set to 0.35 and updated quarterly using accumulated observation data during subsequent monitoring. A sufficient fire correlation is considered to exist between two nodes only when the overlap coefficient is greater than the preset overlap threshold, and a connection is established only at this point.

[0055] The direction of directed spatial adjacency edges points from nodes with higher expected prior probabilities to nodes with lower expected probabilities, simulating the physical trend of fire spreading from high-risk to low-risk areas, consistent with the actual laws of fire propagation. This direction is independent of the node's geographical location, depending only on the comparison of the expected values ​​of the prior probability distributions. For example, if the expected prior probability of a fire in grid A is greater than that in grid B, then a directed edge is established from A to B. The edge weights of the directed spatial adjacency edges are set as overlap coefficients, so that the edge weights characterize the ease or probability of fire propagation in space; the larger the overlap coefficient, the larger the edge weight, indicating a stronger correlation in fire propagation.

[0056] The fire intensity prediction sample sequence is the fire intensity prediction value generated by Monte Carlo ensemble simulation for the same grid node at consecutive time steps. By calculating the Pearson correlation coefficient of the fire intensity prediction sample sequence and taking the absolute value as the temporal autocorrelation coefficient, the linear correlation between the fire status of the same grid node at adjacent time steps can be measured. The Pearson correlation coefficient can capture the evolutionary inertia of the fire in the time dimension, that is, the correlation between the fire status at the current moment and the fire status at the previous moment. Taking the absolute value is to represent the correlation strength, without distinguishing between positive and negative correlation. For example, if the fire intensity prediction sample sequences of a grid node are highly correlated in two consecutive time steps, it indicates that the fire status of that node has strong continuity. Establishing temporal self-connecting edges between adjacent time steps and setting the edge weights to the temporal autocorrelation coefficient allows the dynamic spatiotemporal graph to automatically adjust the temporal constraint strength between nodes according to the continuity of fire evolution, thereby more accurately representing the dynamic evolution of the fire in time.

[0057] As can be seen from the above, this application first perturbs the input parameters of the physical fire behavior model through Monte Carlo ensemble simulation to obtain the predicted mean and variance of the fire line intensity of each grid, fully covering the prediction fluctuations caused by input uncertainty; then, after normalizing the statistical results, a prior probability in the form of a beta distribution is constructed to adapt to the bounded probability representation of the fire state; then, a directed spatial adjacency edge is established based on the overlap coefficient of the prior distribution of adjacent nodes, with the weight linked to the overlap coefficient, to quantify the correlation strength and diffusion direction of the fire state between nodes; at the same time, a time-series self-connected edge is constructed based on the Pearson correlation coefficient of the predicted samples of adjacent time steps to characterize the state continuity in the time dimension.

[0058] In one embodiment of this application, a space-air-ground forest fire monitoring method based on Bayesian variational inference further includes: Obtain the meteorological forecast wind field at the corresponding moment in the dynamic spatiotemporal map, and determine the prevailing wind direction based on the meteorological forecast wind field; For each pair of spatially adjacent grid nodes, calculate the angle between the azimuth from the current grid node to the adjacent grid node and the prevailing wind direction; If the included angle is less than the preset diffusion angle threshold, then the first overlap threshold is used as the preset overlap threshold. If the included angle is greater than or equal to the preset diffusion angle threshold, then the second overlap threshold is used as the preset overlap threshold. The first overlap threshold is less than the second overlap threshold.

[0059] In this embodiment, the weather forecast wind field is a data set describing the distribution of wind speed and direction within a specific area and time period. It can be acquired in various ways, such as real-time collection through a distributed weather sensor network deployed in forest areas, or by obtaining output data from numerical weather prediction models provided by regional meteorological departments. Determining the dominant wind direction refers to extracting the primary wind direction that has the greatest impact on fire spread from complex wind field data. This is achieved through statistical analysis of wind field vector data, for example, by calculating the mode of wind direction frequency distribution within a certain time window, or by using a weighted average wind direction vector.

[0060] Spatially adjacent grid nodes refer to two grid cells that are geographically adjacent to each other in a dynamic spatiotemporal map. The azimuth angle from the current grid node to the adjacent grid node is calculated geometrically using the geographic or projected coordinates of these two grid nodes. For example, the relative azimuth angle can be calculated using the difference in latitude and longitude or the difference in planar coordinates between two points, and the arctangent function can be used. The prevailing wind direction angle refers to the angular difference between the calculated inter-grid azimuth angle and the determined prevailing wind direction. This prevailing wind direction angle is calculated using the geometric principle of vector dot product.

[0061] The diffusion angle threshold is a pre-set angle value used to define the effective range of fire spread under the influence of the prevailing wind direction. This diffusion angle threshold is statistically set based on typical forest fire patterns. Principal component analysis is performed on historical satellite fire images to extract the angle between the long axis of the fire and the wind direction. The 85th percentile of the angle distribution of all samples is taken as the diffusion angle threshold, with an initial value of 45°. It is recalculated every six months based on newly added fire data.

[0062] The overlap threshold is a standard used to determine whether a spatial adjacency edge is established between two grid nodes. The first overlap threshold and the second overlap threshold are two different values, where the first overlap threshold is less than the second overlap threshold. For example, the first overlap threshold is set to the baseline overlap threshold (i.e., the aforementioned preset overlap threshold) multiplied by a direction coefficient of 0.7, and the second overlap threshold is set to the baseline overlap threshold multiplied by a direction coefficient of 1.3, to reflect the principle that downwind connections are easier to establish and upwind connections are more difficult to establish.

[0063] As can be seen from the above, this application first determines the dominant wind direction in the forest area based on meteorological forecast wind fields, clarifying the advantageous direction of fire spread; then, it adopts differentiated overlap thresholds for spatially adjacent nodes in different directions: the downwind area has a high probability of fire spread, so the first overlap threshold is used, making it easier to establish spatial adjacent edges and strengthen the correlation transmission of the spread direction; the upwind area has a low probability of fire spread, so the second overlap threshold is used to strictly control the establishment of connections and weaken redundant connections in non-dominant directions. This embodiment makes the spatial connection of the graph structure anisotropic, which can match the wind direction driving characteristics of real fire spread, avoid the computational redundancy and inference interference introduced by irrelevant edges, and further improve the physical reality of the prior topology and the computational efficiency of graph inference.

[0064] In one embodiment of this application, a dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node are input into a preset graph network based on factor graphs and expectation propagation. The posterior probability distribution characterizing the fire status of each grid node is then calculated using Bayesian variational inference, including: The prior probability distribution of each grid node is constructed as a prior factor, the modal data in the multimodal feature vector is constructed as an observation factor based on a preset likelihood function, and the edge weights carried by spatial adjacency edges and temporal self-connection edges in the dynamic spatiotemporal graph are converted into smoothing factors between adjacent nodes. The prior factor, observed factor, and smoothing factor are respectively attached to the corresponding raster nodes and node pairs to generate a factor map. The expectation propagation algorithm is used to perform message iteration on the factor graph: an approximate posterior distribution belonging to the beta distribution family is maintained for each node; In each iteration, for the prior factor and observed factor attached to a single node, the factor currently being processed is denoted as the target factor. The current message of the target factor is removed from the current approximate posterior distribution of the node to obtain the cavity distribution. The potential function carried by the target factor is combined to obtain the skew distribution. The skew distribution is projected onto the beta distribution family through moment matching to obtain the updated approximate posterior distribution. The message of the target factor passed to the variable node is then updated. When the changes in the approximate posterior distribution parameters of all nodes are less than the preset threshold, the iteration terminates, and the current approximate posterior distribution of each node is output as the posterior probability distribution.

[0065] In this embodiment, the prior factor is a component in the probabilistic graphical model used to introduce prior knowledge, representing the initial belief about the state of a certain variable before any observation data is available. In this embodiment, this prior knowledge originates from the prior probability distribution of the fire state of each grid node generated by the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model. By directly using the parameters of the beta distribution (shape parameters α and β) as the parameters of the prior factor, whose potential function is proportional to the probability density function of the beta distribution, the prediction of the fire state by the physical model is directly encoded into the inference framework.

[0066] In this embodiment, each modal data in the multimodal feature vector is constructed as an observation factor based on a preset likelihood function. The observation factor is used to link actual observed data with variable states, and the likelihood function describes the probability of observed data occurring under a given variable state. For example, for a binary fire situation (fire present / no fire), a Bernoulli likelihood model is used to convert various modal data, such as satellite hotspot monitoring data, UAV inspection video stream data, and ground IoT sensor time-series data, into likelihood intensity values ​​of fire occurrence, which serve as the potential function of the observation factor.

[0067] The smoothing factor in this embodiment describes the dependency between adjacent variables (grid nodes) to encourage adjacent variables to have similar states or to characterize the physical association between them, while the edge weights quantify the strength of this association. For example, the potential function of the smoothing factor is designed to penalize the difference in state between adjacent nodes. The greater the difference in the fire state between two adjacent nodes, the smaller the potential function value, and vice versa. The edge weights, as a parameter in the potential function, adjust the smoothing intensity.

[0068] In this embodiment, the factor graph network adopts a hierarchical architecture, specifically including three layers: node-level factor layer, edge-level factor layer, and global coordination layer. The node-level factor layer maintains a variable node xi for each grid node, representing the fire status of the node, with a value of a continuous value in the range of 0 and 1, as well as the prior factor fprior(xi) and the observation factor fobs(xi) attached to the node.

[0069] The edge-level factor layer establishes a smoothing factor fsmooth(xi,xj) for each spatially adjacent edge and maintains a temporally self-connected edge. The potential function of this factor is defined as: ψ(xi,xj=exp(-wij·(xi-xj)). 2 / (2σs 2 )), where xi and xj represent the fire status variables of any two variable nodes (grid nodes) linked by spatial adjacency edges or temporal self-connection edges in the dynamic spatiotemporal graph, wij is the weight of the connection edge, and σs is the smoothing scale parameter, which is preset to 0.15. This potential function encourages adjacent nodes to maintain continuity in fire status.

[0070] The global coordination layer contains a superfactor fglobal({xi}), whose potential function is the spatial second derivative regularization term of the fire status of each node, used to suppress isolated noise points. Its expression is: ψglobal=exp(-λg·Σi(∇²xi) 2 ), where λg is the global regularization coefficient, preset to 0.05, and ∇²xi is the spatial Laplace operator for the fire state at node i.

[0071] The input data stream of the factor graph network consists of: the multimodal feature vectors of each grid node, the beta distribution parameters α and β of the prior probability distribution of each node, and the adjacency weight matrix W, which is composed of the weights of all directed spatial adjacency edges and temporal self-connection edges in the dynamic spatiotemporal graph. The output of the factor graph network is the approximate posterior distribution of each variable node, in beta distribution form.

[0072] This embodiment generates a factor graph by attaching prior factors, observation factors, and smoothing factors to the corresponding raster nodes and node pairs, respectively, thereby unifying all prior knowledge, observation data, and spatiotemporal correlations into a graphical model.

[0073] In this embodiment, the expectation propagation algorithm is used to iterate messages on the factor graph, maintaining an approximate posterior distribution belonging to the beta distribution family for each node. In each iteration, for the prior factors and observed factors attached to a single node, the currently processed factor is denoted as the target factor. First, the current message of the target factor is removed from the current approximate posterior distribution of that node to obtain the cavity distribution. The cavity distribution represents the beliefs of all factors except the target factor regarding the variable. Next, the potential function carried by the target factor is combined to obtain the skewed distribution, which contains the latest information of the target factor regarding the variable. Since the skewed distribution may not belong to the beta distribution family, it needs to be projected onto the beta distribution family through moment matching to obtain an updated approximate posterior distribution and update the message passed by the target factor to the variable node. Moment matching involves calculating the mean and variance of the skewed distribution and then solving for the shape parameters of the beta distribution.

[0074] The iteration terminates when the changes in the approximate posterior distribution parameters of all nodes are less than a preset threshold, i.e., the sum of the absolute changes in the beta distribution shape parameters α and β is less than the preset threshold. The current approximate posterior distribution of each node is then output as a posterior probability distribution. This preset threshold is determined as follows: During the initialization phase, the expectation propagation algorithm is run using simulation data. The changes in the posterior parameters of each node are recorded after each iteration. The median of the parameter changes for all nodes at the 10th iteration (basic convergence) is taken as the convergence threshold, set to 0.01. This termination condition ensures that the algorithm stops after reaching sufficient accuracy, avoiding unnecessary calculations and outputting the final posterior probability distribution of the fire situation.

[0075] As can be seen from the above, this application first transforms prior probabilities, multimodal observations, and spatiotemporal connectivity constraints into prior factors, observation factors, and smoothing factors, respectively, and maps them uniformly into a factor graph framework to construct a standardized probabilistic inference carrier. Then, it uses the expectation propagation algorithm to perform message iteration, maintaining an approximate posterior for each node's beta distribution family by updating the posterior parameters. Finally, it terminates the iteration with the parameter change as the convergence condition, outputting a stable posterior probability distribution. This embodiment achieves efficient fusion of multi-source information while ensuring a unified probability distribution form, balancing the inference accuracy and computational efficiency of large-scale grid nodes.

[0076] In one embodiment of this application, message iteration on the factor graph using the expectation propagation algorithm further includes: For the smoothing factor attached to the node pair, the current message passed to each node by the smoothing factor is removed from the current approximate posterior distribution of each of the two nodes in the node pair to obtain the cavity distribution of each node. After multiplying the two cavity distributions, and then multiplying by the potential function of the smoothing factor, the joint tilt distribution of the node with respect to the fire situation is obtained. Calculate the marginal moments of the two variables under the joint skewed distribution, project each marginal distribution onto the beta distribution family through moment matching, obtain the approximate posterior distribution updated by each of the two nodes, and update the smoothing factor of the message passed to the two nodes respectively.

[0077] In this embodiment, the smoothing factor is a technical feature used to characterize the interrelationships or constraints between node pairs in a dynamic spatiotemporal graph. Its function is to ensure that the fire state estimates of adjacent nodes (whether spatially adjacent or temporally self-connected) maintain consistency and smoothness, thereby more accurately representing the physical laws of fire spread. In this embodiment, the current message passed to each node by the smoothing factor is removed to eliminate the influence of the smoothing factor on the node's approximate posterior distribution in the previous iteration, thus obtaining the cavity distribution of that node. The cavity distribution represents the combined belief of all other factors regarding the node's state, excluding the currently processed smoothing factor. The purpose of this step is to avoid information being repeatedly calculated during iteration, ensuring that the current smoothing factor can be updated based on node states unaffected by its previous influence. Specifically, if both the approximate posterior distribution and the message belong to the exponential family distribution, the cavity distribution is obtained by subtracting the natural parameters of the corresponding message from the natural parameters of the approximate posterior distribution.

[0078] Multiplying the two cavity distributions together, and then multiplying by the potential function of the smoothing factor, yields the joint tilted distribution of the fire state for the node pair. This step combines the independent cavity distributions of two interconnected nodes with the potential function defined by the smoothing factor between them to construct an unnormalized joint probability distribution, i.e., the joint tilted distribution. This joint tilted distribution integrates the individual cavity distributions of the two nodes and the interaction relationship between them characterized by the smoothing factor. Through this multiplication operation, the physical association between the node pairs can be directly incorporated into the joint probability model, thus providing a more comprehensive description of the fire state of the node pairs. For example, the potential function is a similarity measure based on the beta distribution, encouraging two nodes to have similar probabilities of fire occurrence.

[0079] After obtaining the joint tilt distribution of nodes in relation to the fire situation, it is necessary to extract individual information for each node and update it to their respective approximate posterior distributions. First, the joint tilt distribution is marginalized to obtain the marginal distributions for each node. Next, the moments (e.g., mean and variance) of these marginal distributions are calculated. This implementation uses a moment matching method to match the calculated marginal moments with the moments of the beta distribution, thereby determining the shape parameters α and β of the beta distribution, and thus obtaining the updated approximate posterior distribution for each node.

[0080] As can be seen from the above, this application addresses the smoothing factor between node pairs by first removing historical messages of the smoothing factor from the current posterior of both nodes to obtain their respective cavity distributions, ensuring unbiased message transmission. Then, the two cavity distributions are multiplied and superimposed with the potential function of the smoothing factor to obtain the joint tilt distribution of the fire status of the node pair, completing the fusion of spatiotemporal constraints. Next, the marginal moments of the two variables are calculated separately, and the results are projected back to the beta distribution family through moment matching, synchronously updating the approximate posterior and corresponding messages of the two nodes. This process effectively injects the constraints of spatial adjacency and temporal continuity into the posterior estimate of each node, smoothing the random fluctuations of single-node observations, strengthening the spatiotemporal consistency of fire status estimation in contiguous areas, and improving the coherence and stability of regional fire situation projection.

[0081] In one embodiment of this application, the modal data in the multimodal feature vector are constructed as observation factors based on a preset likelihood function, including: For the same grid node, the likelihood of a fire is calculated using the Bernoulli likelihood model based on satellite hotspot monitoring data, UAV inspection video stream data, and ground IoT sensor time series data, respectively. Calculate the difference between each pair of the three sets of likelihood intensity values. If the difference between any two modal likelihood intensity values ​​is greater than the preset modal consistency threshold, obtain the false alarm rate and false negative rate of each modal data within the preset historical time window, and use the normalized false alarm rate and false negative rate to calculate the confidence weight of each modality. The joint likelihood intensity is generated by weighting the modal likelihood intensity values ​​based on the confidence weight, and a joint observation factor is constructed based on the joint likelihood intensity. If the difference between the likelihood values ​​of any two modes is less than or equal to the preset modal consistency threshold, the arithmetic mean of the likelihood values ​​of each mode is directly calculated, and the observation factor is constructed based on the arithmetic mean.

[0082] In this embodiment, the Bernoulli likelihood model is a special case of the binomial distribution, suitable for describing events with only two possible outcomes (such as a fire occurring or not occurring). This Bernoulli likelihood model maps heterogeneous information such as satellite hotspot monitoring data, UAV inspection video stream data, and time-series data from ground-based IoT sensors to likelihood values ​​between 0 and 1, where a higher value indicates a greater probability of a fire occurring.

[0083] For example, for satellite hotspot monitoring data, based on the anomalies in radiance or temperature in the infrared band, a preset sigmoid function is used to convert them into a likelihood value indicating the occurrence of a fire. For UAV inspection video stream data, a deep learning model is used to analyze the video frames, identify fire features such as flames and smoke, and output the corresponding confidence score, which serves as the likelihood value indicating the occurrence of a fire.

[0084] In multi-source data fusion, different sensors or monitoring methods may produce inconsistent observations of the same event due to their own characteristics, environmental interference, or malfunctions. This inconsistency is quantified by calculating the difference between pairwise modal likelihood values. For example, the difference is represented by calculating the relative difference between two modal likelihood values, such as (|likelihood value A - likelihood value B|) / (max(likelihood value A, likelihood value B) + q), where q is a small positive number to avoid division by zero.

[0085] If the difference in likelihood intensity values ​​between any two modalities exceeds a preset modality consistency threshold, a simple averaging is no longer performed blindly. Instead, the false alarm rate and false negative rate of each modality's data within a preset historical time window are obtained, and the credibility weight of each modality is calculated using the normalized false alarm rate and false negative rate. When conflicts are detected between different modalities, a dynamic weighting mechanism based on historical performance is adopted. The false alarm rate refers to the probability of misreporting a non-fire situation as a fire, and the false negative rate refers to the probability of failing to detect a real fire. The false alarm rate and false negative rate characterize the inherent reliability of each modality's data source. By statistically analyzing these indicators over a 7-day period and normalizing them, credibility weights are obtained for subsequent weighted fusion, thereby reducing the impact of unreliable data sources on the final judgment. For example, the credibility weight is calculated based on the reciprocal of the false alarm rate and the false negative rate or a combination thereof, such as weight = 1 / (false alarm rate + false negative rate + ε), where ε is a small constant. In this embodiment, normalization is achieved by setting the sum of the weights of all modalities to 1. Alternatively, the credibility weights can be calculated using a Bayesian method, taking historical false alarm rates and false negative rates as prior information, combined with current observation data, to calculate the posterior reliability of each modality data source, and then using this reliability as the weights.

[0086] A weighted geometric mean is calculated based on the confidence weights of the likelihood values ​​of each mode to generate a joint likelihood strength, and a joint observation factor is constructed based on the joint likelihood strength. When there are conflicts between modal data, the weighted geometric mean is a more robust fusion strategy. The geometric mean is more sensitive to extreme values, especially those close to zero, and can effectively suppress the negative impact of low-confidence modes while amplifying the contribution of high-confidence modes. The joint likelihood strength obtained in this way can more accurately represent the true probability of a fire occurring, and a more robust joint observation factor can be constructed based on this for Bayesian variational inference in the factor graph. The observation factor in the factor graph represents the influence of the observed data on the fire status. For example, the weighted geometric mean is expressed as (likelihood strength value 1^weight 1 × likelihood strength value 2^weight 2 × ... × likelihood strength value N^weight N)^(1 / Σweight). Then, this joint likelihood strength is used as part of the potential function to construct the joint observation factor.

[0087] The preset modal consistency threshold was determined by the following method: Synchronous observation data from three types of sensors were collected in the forest area under both fire-free and fire-affected conditions. The absolute difference in the pairwise modal likelihood intensity values ​​under each condition was calculated. Kernel density estimation was used to fit the distribution of the difference values. Under the premise that the probability of the difference value being less than the threshold (i.e., not triggering a conflict) under the fire-affected condition is not less than 95%, the threshold that minimizes the probability of false conflict triggering under the fire-free condition was selected as the modal consistency threshold. The initial value was set to 0.25, and it was recalculated and updated every 100 newly calibrated samples.

[0088] If the difference in likelihood values ​​between any two modes is less than or equal to a preset modal consistency threshold, it indicates a high degree of consistency among the modal data. In this case, the arithmetic mean of the likelihood values ​​of each mode is directly calculated, and observation factors are constructed based on the arithmetic mean. When the consistency between different modal data is high, i.e., the difference does not reach the preset modal consistency threshold, it indicates that the observation results of each data source on the fire situation are basically consistent. In this case, the arithmetic mean is used for fusion. The arithmetic mean can make full use of the complementary information of each modal data and provide a comprehensive judgment when the data quality is reliable. Constructing observation factors based on the arithmetic mean can simplify the calculation process while maintaining the accuracy of the fusion results. For example, the arithmetic mean is calculated as (likelihood value 1 + likelihood value 2 + ... + likelihood value N) / N; then, this arithmetic mean is used as part of the potential function to construct observation factors.

[0089] As can be seen from the above, this application calculates the fire likelihood intensity based on the Bernoulli likelihood model for three types of modal data, achieving a probabilistic unified representation of multi-source observations. Then, it judges the consistency of observations by the difference between modalities: when there is no conflict, an arithmetic mean is used to simplify the calculation process and improve fusion efficiency; when conflicts occur, a confidence weight is calculated based on the historical false alarm rate and missed detection rate of each modality, and a weighted geometric mean is used to generate the joint likelihood. This adaptive fusion mechanism can ensure lightweight computation when multi-source data is consistent, and suppress the interference of low-confidence modalities when data conflicts occur, thus preserving the complementary advantages of multi-source data and improving the robustness of the observation input.

[0090] In one embodiment of this application, when the difference in likelihood intensity values ​​between any two modes is greater than a preset modal consistency threshold, the method further includes: At the current iteration step, obtain the expected posterior probability distribution of the neighboring grid nodes connected to the current grid node through spatial adjacency edges at the previous time step, as the spatial neighborhood posterior expectation, and the expected posterior probability distribution of the current grid node at the previous time step, as the temporal posterior expectation. For each mode, calculate the first deviation of the mode likelihood intensity value from the spatial neighborhood posterior expectation, and the second deviation from the temporal posterior expectation; If the first deviation of a certain mode is greater than the preset spatial deviation threshold, and the second deviation is also greater than the preset temporal deviation threshold, then the mode is marked as a spatiotemporal anomalous mode. When generating joint observation factors, likelihood intensity values ​​marked as spatiotemporal anomalous modes are removed; For the likelihood values ​​of the remaining unlabeled modes, a weighted geometric mean is calculated based on the confidence weight to generate a joint likelihood value, and a joint observation factor is constructed based on the joint likelihood value.

[0091] In this embodiment, the spatial neighborhood posterior expectation represents the fire status estimate of the area surrounding the current grid node at the previous time step, providing a dynamic reference for assessing the spatial consistency of the current grid node's observation data. This expectation is obtained by averaging the expectations of the posterior probability distributions of all neighboring grid nodes at the previous time step. The temporal posterior expectation represents the historical evolution trend of the fire status of the current grid node itself, providing a dynamic benchmark for assessing the temporal consistency of the current observation data. This temporal posterior expectation is taken as the expectation of the posterior probability distribution at the previous time step.

[0092] In this embodiment, the first and second biases are used to quantify the degree of deviation between the current modal observation data and the spatiotemporal reference. The first bias is calculated as the absolute difference between the modal likelihood intensity value and the posterior expectation of the spatial neighborhood. The second bias is calculated in a similar manner to the first bias, that is, as the absolute difference between the modal likelihood intensity value and the temporal posterior expectation.

[0093] If the first deviation of a certain mode is greater than a preset spatial deviation threshold, and the second deviation is simultaneously greater than a preset temporal deviation threshold, then the mode is marked as a spatiotemporally anomalous mode. The preset spatial and temporal deviation thresholds are calculated offline based on historical normal operating data: multimodal observation data of the grid node and its neighboring nodes over the past 30 days under conditions without abnormal fires are extracted, and the absolute values ​​of the deviations between the likelihood intensity value of each mode and the posterior expectation of the spatial neighborhood and the posterior expectation of the temporal sequence are calculated. The 95th quantile of each distribution is taken as the initial threshold. Subsequently, the threshold is updated daily, and an exponentially weighted moving average is used to smooth the threshold changes, with a smoothing factor of 0.2 to balance stability and adaptability.

[0094] For the likelihood values ​​of the remaining unlabeled modes, a weighted geometric mean is calculated based on the confidence level to generate a joint likelihood strength. A joint observation factor is then constructed based on this joint likelihood strength. After removing outlier modes, reliable mode data that has undergone spatiotemporal consistency verification is used for fusion. The weighted geometric mean comprehensively considers the confidence level of each mode data, generating a more robust and accurate joint likelihood strength. The joint observation factor constructed based on this joint likelihood strength will serve as observation information in the factor graph, providing high-quality input for Bayesian variational inference.

[0095] As can be seen from the above, in scenarios where multimodal observations conflict, this application calculates the double deviation between the likelihood values ​​of each mode and the spatiotemporal reference, marks and removes anomalous modes that simultaneously deviate from both spatial correlation and temporal continuity patterns, and filters out invalid observations caused by equipment failures and sudden noise. Then, a weighted geometric average is performed on the remaining valid modes based on their credibility weights to generate the final joint observation factor. This embodiment adds spatiotemporal consistency verification to the credibility weighting, which can eliminate abnormal data interference from the perspective of physical evolution laws, further reducing the probability of false detections and false alarms, and improving the robustness of observation inputs and the reliability of posterior inference under extreme conditions.

[0096] In one embodiment of this application, adjusting the graph structure of a dynamic spatiotemporal graph based on the degree of cognitive conflict includes: Grid nodes with cognitive conflict degree greater than the preset cognitive conflict threshold are marked as nodes to be reconstructed. All spatial adjacency edges connected to the nodes to be reconstructed in the dynamic spatiotemporal graph are removed. At the same time, the weight of the temporal self-connection edge corresponding to the node to be reconstructed is multiplied by the decay coefficient corresponding to the node to be reconstructed. For each node to be reconstructed, calculate the Helling distance between its posterior probability distribution and that of its spatially adjacent nodes. When the Hellinger distance is greater than the reconnection threshold, a directed spatial adjacency edge is established between the node to be reconstructed and the adjacent node, and the edge weight is set to the Hellinger distance. For nodes whose cognitive conflict degree is less than or equal to the preset cognitive conflict threshold, their original spatial adjacency edges and temporal self-connection edges are retained, and the edge weights of the spatial adjacency edges and temporal self-connection edges are updated to the product of the current edge weight and the decay coefficient; the decay coefficient is 1 minus the ratio of the cognitive conflict degree to the preset maximum KL divergence.

[0097] In this embodiment, cognitive conflict is an indicator that measures the degree of difference between model predictions and actual observations. Specifically, it measures the difference between the model's prior probability distribution and the posterior probability distribution obtained through Bayesian variational inference. This embodiment obtains this value by calculating the node-wise KL divergence between the posterior and prior probability distributions. KL divergence is an asymmetric measure used to quantify the difference between two probability distributions. A larger KL divergence value indicates a greater difference between the two distributions, i.e., a greater cognitive conflict. This KL divergence can intuitively characterize the inconsistency between the model's perception of the current fire situation and the actual data.

[0098] The preset cognitive conflict threshold is a pre-defined value used to determine whether the degree of cognitive conflict has reached a point where graph structure adjustment is necessary. This preset cognitive conflict threshold is determined as follows: During initialization, Bayesian inference is run using simulated fire data (including known real fire conditions), and the distribution of KL divergence for all grid nodes is statistically analyzed. The quantile that minimizes the mean square error of the fire condition estimation caused by the reconstructed graph structure is taken as the preset cognitive conflict threshold. After each complete closed-loop iteration, the KL divergence of all nodes in this iteration is sorted, and the 75th percentile is taken as the cognitive conflict threshold for the next iteration. The threshold is limited to a minimum of 0.1 and a maximum of 2.0 to avoid frequent reconstruction due to an excessively small threshold or reconstruction delays due to an excessively large threshold.

[0099] Nodes to be reconstructed refer to grid nodes whose cognitive conflict level exceeds a preset cognitive conflict threshold. These nodes represent areas where there is a significant discrepancy between the model's fire situation prediction and actual observation, indicating a sudden change in the fire situation in that area or a serious deficiency in the model's understanding of that area. This embodiment marks these nodes to be reconstructed in order to perform targeted graph structure adjustments to correct the model's cognitive biases in these areas.

[0100] Removing all spatial adjacency edges connected to the node to be reconstructed in the dynamic spatiotemporal graph aims to sever any existing spatial connections between the node and its surrounding nodes that may have become invalid or inaccurate. When a node has a high level of cognitive conflict, it means that the relationship between it and its neighboring nodes regarding the spread or impact of the fire may have changed, and the original spatial adjacency edges can no longer accurately represent the current fire situation. Removing these edges can prevent the propagation of erroneous information in the graph network and prepare for the subsequent establishment of more accurate connections.

[0101] For example, this can be achieved by deleting the corresponding edge records in the graph data structure, or by setting the weights of these edges to zero or a minimum value. Simultaneously, the weights of the temporal self-connected edges corresponding to the node to be reconstructed are multiplied by the decay coefficient corresponding to the node. Temporal self-connected edges reflect the node's own state continuity over time. When a node is marked as needing reconstruction, it indicates that its state may have changed, and the original assumption of temporal continuity may no longer hold. By multiplying by the decay coefficient, the node's dependence on its historical state over time is reduced, making it more susceptible to the influence of new observation data and model updates, thereby accelerating the model's adaptation to abrupt changes. The decay coefficient is a positive number less than 1, such as 0.5, or it can be dynamically calculated based on the degree of cognitive conflict.

[0102] Hellinger distance is another measure of the similarity between two probability distributions. Unlike KL divergence, Hellinger distance is symmetric, and its value lies between 0 and 1, where 0 indicates that the two distributions are completely identical, and 1 indicates that the two distributions do not overlap at all. In this embodiment, the Hellinger distance between the posterior probability distributions of the node to be reconstructed and its spatial neighbors is calculated to quantify the degree of difference in fire status between these nodes. For example, the more similar the posterior probability distributions of two nodes, the smaller the Hellinger distance, indicating that their fire statuses are closer and there is a stronger correlation.

[0103] The reconnection threshold is used to determine whether spatial adjacency edges need to be re-established between the node to be reconstructed and its adjacent nodes. When the calculated Hellinger distance is greater than the reconnection threshold, it indicates that there is a difference in the fire situation between the two nodes. However, this difference is precisely the manifestation of fire spread or diffusion, so a new connection needs to be established to capture this dynamic relationship. The reconnection threshold is determined as follows: based on the logistic regression model between the Hellinger distance of adjacent nodes in historical fire spread data and the actual fire propagation probability, with the goal of maximizing the model's prediction accuracy, a grid search is used, and the optimal threshold is selected within the interval [0.1, 0.8] with a search step size of 0.05. The initial default value is set to 0.35. Subsequently, after processing every 100 nodes to be reconstructed, the prediction performance feedback of the reconstructed graph structure is collected, and the threshold is fine-tuned using an online gradient descent method with a step size of 0.01.

[0104] This embodiment establishes directed spatial adjacency edges from the node to be reconstructed to its adjacent node, and sets the edge weights to Helling distance. This operation aims to reconstruct spatial relationships that better reflect the current fire situation based on the actual differences in fire conditions between nodes. Directed edges can better simulate the directionality of fire spread, such as from areas with more severe fires to affected areas. Using Helling distance as the edge weight ensures that nodes with greater differences in fire conditions have stronger connections, thus better reflecting the potential paths of fire spread in the graph network. For example, this can be achieved by adding new edge records to the graph data structure and assigning them the calculated Helling distance as weights.

[0105] For nodes with a cognitive conflict level less than or equal to a preset cognitive conflict threshold, their original spatial adjacency edges and temporal self-connected edges are retained. This indicates that the model's prediction of the fire situation for these nodes is relatively accurate, and their original graph structure connections remain effective. Therefore, large-scale structural reconstruction is unnecessary; instead, the existing spatial adjacency edges and temporal self-connected edges are retained to maintain the stability and continuity of the graph network. Simultaneously, the edge weights of both spatial adjacency edges and temporal self-connected edges are updated to the product of the current edge weight and a decay coefficient. Even if the cognitive conflict level does not exceed the threshold, the existence of some cognitive conflict implies that the model's perception of these nodes may have slight biases. By multiplying the edge weights by the decay coefficient, the association strength of these nodes is fine-tuned, allowing it to adaptively adjust according to the real-time cognitive conflict level. This smooth weight update mechanism enables the model to respond to subtle changes in the fire situation's evolution while maintaining overall stability. The attenuation coefficient is 1 minus the ratio of cognitive conflict level to the preset maximum KL divergence. The calculation method of the attenuation coefficient reflects its negative correlation with cognitive conflict level: the higher the cognitive conflict level, the smaller the attenuation coefficient, and thus the stronger the attenuation effect on edge weights; conversely, the lower the cognitive conflict level, the closer the attenuation coefficient is to 1, and the weaker the attenuation effect on edge weights. The preset maximum KL divergence is used to normalize the cognitive conflict level, ensuring that the attenuation coefficient varies within a reasonable range. For example, it can be set to the historically observed maximum KL divergence value or an empirical upper limit.

[0106] As can be seen from the above, this application first uses cognitive conflict level as the criterion, marking high-conflict nodes as nodes to be reconstructed, removing their original spatial adjacency edges and attenuating the weights of temporal edges, thus breaking the constraints of the failed prior topology; then, based on the Hellinger distance of the posterior distribution, it remeasures the state associations between nodes, and reconstructs directed spatial edges when a threshold is met, allowing the topology to match the current real fire situation; for low-conflict nodes, the original connections are retained and the weights are smoothly attenuated, taking into account structural stability. This adaptive graph reconstruction mechanism can dynamically correct the spatiotemporal graph structure as the fire evolves, continuously eliminating cognitive biases between the prior model and actual observations, avoiding inference failures caused by fixed topology, and ensuring the long-term effectiveness of Bayesian variational inference during the dynamic spread of the fire.

[0107] In one embodiment of this application, establishing a directed spatial adjacency edge between the node to be reconstructed and the adjacent node, from the node to be reconstructed to the adjacent node, includes: Obtain the meteorological forecast wind field at the corresponding moment in the dynamic spatiotemporal map, and determine the prevailing wind direction based on the meteorological forecast wind field; For the node to be reconstructed and any of its spatially adjacent nodes, calculate the azimuth angle of the spatially adjacent node relative to the node to be reconstructed. If the angle between the azimuth and the prevailing wind direction is within the preset downwind angle range, the first reconnection threshold is used, and a directed spatial adjacency edge is established from the node to be reconstructed to the adjacent node in the space. If the angle between the azimuth and the prevailing wind direction is within the preset upwind angle range, the second reconnection threshold is used, and a directed spatial adjacency edge is established from the adjacent node in the space to the node to be reconstructed. The first reconnection threshold is less than the second reconnection threshold.

[0108] In this embodiment, for the node to be reconstructed and any of its spatially adjacent nodes, the azimuth angle of the spatially adjacent node relative to the node to be reconstructed is calculated to determine the relative position of the adjacent node around the node to be reconstructed. In a geographic coordinate system, the azimuth angle is calculated using the latitude and longitude coordinates of the two nodes through plane geometric formulas, such as using the ATAN2 function combined with the coordinate difference between the two points. If the raster node uses a Cartesian coordinate system, the relative angle is directly calculated using the difference between the coordinates of the two points, such as using the arctangent function.

[0109] If the angle between the azimuth and the prevailing wind direction is within the preset downwind angle range, the first reconnection threshold is applied, and directed spatial adjacency edges are established from the node to be reconstructed to its spatially adjacent nodes. The downwind angle range refers to a certain angular interval extending to both sides of the prevailing wind direction, representing the direction in which fire is more likely to spread. The preset downwind angle range is set, for example, ±45 degrees or ±60 degrees of the prevailing wind direction.

[0110] If the angle between the azimuth and the prevailing wind direction is within the preset upwind angle range, a second reconnection threshold is applied, and directed spatial adjacency edges are established from adjacent nodes in that space to the node to be reconstructed. The upwind angle range refers to the area opposite to or significantly deviating from the prevailing wind direction, representing a direction where fire is less likely to spread. The preset upwind angle range is set, for example, from ±90 degrees to ±180 degrees of the prevailing wind direction (i.e., the headwind direction).

[0111] The first and second reconnection thresholds are obtained by multiplying the baseline reconnection threshold (i.e., the reconnection threshold mentioned above) by a direction adjustment coefficient: the downwind adjustment coefficient is 0.6, and the upwind adjustment coefficient is 1.4. The adjustment coefficients are determined through statistical analysis of historical fire wind direction and spread direction, making the downwind connection establishment more sensitive than the upwind connection.

[0112] As can be seen from the above, this application introduces the dominant wind direction as a physical constraint during the dynamic reconstruction of spatial edges: the downwind direction is the direction in which fire is prone to spread, so a lower first reconnection threshold is adopted, and directed edges in the downwind direction are established to strengthen the correlation and transmission of the spread direction; the upwind direction has a low probability of fire spread, so a higher second reconnection threshold is adopted, and directed edges in the upwind direction are established to strictly control the establishment of reverse connections. This design ensures that the dynamically adjusted graph structure always conforms to the wind-driven characteristics of fire behavior, further improving the physical realism of the spatial topology, reducing the computational overhead caused by invalid connections, and ensuring the rationality of the direction of fire propagation in posterior reasoning, thereby improving the accuracy of reasoning after graph structure iteration.

[0113] In one embodiment of this application, the preset mutation conditions include: The joint risk field is segmented by threshold, and a set of high-risk grids with joint risk values ​​greater than a preset risk threshold is extracted. Several high-risk connected components are then extracted based on spatial adjacency. For each high-risk connected component, for each grid node within that high-risk connected component, calculate the temporal change slope of the model cognitive bias over the most recent K consecutive time steps, and take the average of the absolute values ​​of the temporal change slopes of all grid nodes as the average change slope. Calculate the position difference vector of the geometric center within the high-risk connected domain between the current time step and K time steps ago, and take the absolute value of the projection of the position difference vector onto the unit vector of the prevailing wind direction as the cumulative displacement; If the average slope of change is greater than a preset slope threshold, or the cumulative displacement is greater than a preset displacement threshold, then the high-risk connected region is determined to be a mutation blind zone.

[0114] In this embodiment, the joint risk field is first segmented by thresholding. By segmenting the joint risk field by thresholding, regions with different risk levels can be effectively distinguished. For example, an adaptive thresholding method is used to dynamically determine the threshold based on the statistical distribution of the risk field (such as mean and variance). A preset risk threshold is used to define the minimum risk level of high-risk regions. This preset risk threshold is dynamically determined in each iteration using the Otsu adaptive thresholding method: the joint risk value histogram of all grid nodes in the current joint risk field is statistically analyzed, and the gray level corresponding to the maximum inter-class variance is calculated as the preset risk threshold to achieve optimal separation between high-risk and low-risk regions.

[0115] After extracting a set of high-risk rasters with a joint risk value greater than a preset risk threshold, several high-risk connected components are extracted based on spatial adjacency relationships. Spatial adjacency relationships refer to the spatial proximity of raster nodes, which can be defined using, for example, four-neighborhood or eight-neighborhood. High-risk connected components are continuous regions composed of all spatially adjacent raster nodes in the high-risk raster set, extracted using a depth-first search method.

[0116] The time-varying slope is used to measure the rate of change of model cognitive bias over time. For example, the slope is calculated by performing linear regression analysis on the model cognitive bias over the most recent K consecutive time steps. The number of K time steps is determined based on typical timescales of fire evolution; for example, it can be set to 3, 5, or 10 time steps to capture trends at different time granularities. Then, the average of the absolute values ​​of the time-varying slopes of all grid nodes is taken as the average slope.

[0117] The geometric center refers to the average coordinates of all grid nodes within a high-risk connected component, representing the spatial center of that component. The position difference vector is the vector difference between the geometric center of the high-risk connected component at the current time step and its position K time steps ago, representing the overall spatial movement direction and distance of that connected component.

[0118] The absolute value of the projection of the position difference vector onto the prevailing wind direction unit vector is taken as the cumulative displacement. The prevailing wind direction unit vector refers to the unit vector corresponding to the prevailing wind direction in the forest area at the current moment, obtained through meteorological forecast data or real-time wind field sensor data. Projecting the position difference vector onto the prevailing wind direction unit vector can quantify the degree of fire spread along the prevailing wind direction. Taking the absolute value is to measure the intensity of the displacement, without considering its directionality. The cumulative displacement refers to the actual distance the geometric center of the high-risk connected region has advanced along the prevailing wind direction in the most recent K time steps, characterizing the speed and directionality of fire spread.

[0119] If the average slope change exceeds a preset slope threshold, or the cumulative displacement exceeds a preset displacement threshold, the high-risk connected region is identified as a mutation blind zone. Identifying a mutation blind zone indicates that the fire evolution in that area exhibits high uncertainty, rapid spread, and limited model predictive capabilities, necessitating immediate proactive data collection to obtain more accurate information. The preset slope and displacement thresholds are jointly determined using the following method: Satellite time-series data of all major forest fires that occurred in the target forest area over the past five years are used to extract characteristic samples from the rapid fire expansion phase, i.e., fire line length growth exceeding 50% / hour. The average slope change and cumulative displacement of each characteristic sample at K time steps before the mutation are calculated, and the 10th percentile of these characteristic values ​​is used as the corresponding threshold. If no historical data is available, the slope threshold is set to 0.05 (unit: deviation / minute), and the displacement threshold is set to two grid side lengths.

[0120] As can be seen from the above, this application first performs threshold segmentation on the joint risk field and extracts high-risk connected regions to identify potential risk concentration areas. Then, it quantifies the abrupt change characteristics of the region from two dimensions: first, it calculates the average slope of the model cognitive bias within the region, characterizing the rate of intensity abrupt change in model mismatch; second, it calculates the cumulative displacement of the region's geometric center in the prevailing wind direction, characterizing the degree of drift in the spatial spread of fire. Finally, it identifies abrupt change blind spots. This embodiment identifies model cognitive failure areas from two dimensions: intensity deterioration and spatial drift. It can capture both sudden intensification of fire intensity and directional shifts in fire spread, accurately locating cognitive blind spots in the monitoring system and providing clear targeting basis for proactive supplementary sampling tasks.

[0121] In one embodiment of this application, an active data acquisition instruction is generated based on the mutation blind zone to control a mobile reconnaissance platform to collect supplementary data, including: The grids covered by the mutation blind zone in the joint risk field are mapped to access costs according to the joint risk value, and the areas outside the blind zone are also mapped to access costs according to the risk value. The composite cost map is generated by merging these maps. Obtain the current location, remaining range, and current wind field vector of the mobile reconnaissance platform; For all identified mutation blind zones, the grid node with the largest joint risk value within each mutation blind zone is selected as a candidate target point; the estimated cost of the risky feasible path corresponding to each candidate target point is calculated. When there are candidate paths where the estimated consumption is less than or equal to the remaining endurance and the wind field vector is less than or equal to the safe operation threshold, the feasible blind area with the highest priority index for supplementary mining is selected as the task target. Otherwise, a multi-attribute evaluation matrix is ​​constructed based on the area, risk exposure value, and urgency of replenishment of all mutation blind zones. The replenishment priority index of each blind zone is calculated by the superior-inferior solution distance method, and feasible task objectives are reassigned from high to low based on the replenishment priority index. Among them, the risk exposure value is the sum of the updated joint risk values ​​of all grid nodes in the mutation blind zone. The mobile reconnaissance platform is controlled to perform supplementary data collection along the planned path and the joint risk field is updated in real time within a preset range from the target point. If a new risk intrudes into the planned path, a local replanning is triggered to avoid it.

[0122] In this embodiment, the composite cost map transforms environmental risk information into a quantifiable representation of the travel costs for a mobile reconnaissance platform. It is generated by converting joint risk values ​​into travel costs using a preset mapping function (exponential function); the higher the risk value, the greater the travel cost. For areas outside the blind zone, the risk value is weighted based on factors such as distance from the sudden change blind zone, historical fire data, or the confidence level of current monitoring data, and then mapped to a travel cost.

[0123] The system acquires the current location, remaining battery life, and current wind field vector of the mobile reconnaissance platform. The platform's current location is obtained in real-time via positioning modules such as the Global Positioning System (BeiDou). Remaining battery life is provided by the platform's battery management system. The current wind field vector is measured in real-time by meteorological sensors or obtained from external weather forecast services. For example, the mobile reconnaissance platform carries a small weather station to measure local wind speed and direction in real time; or it receives meteorological data from ground-based weather stations or meteorological satellites via a wireless communication module and combines this data with a terrain model to correct for local wind field variations.

[0124] For all identified mutation blind zones, the grid node with the highest joint risk value within each blind zone is selected as a candidate target point; the estimated cost of the risk-feasible path corresponding to each candidate target point is calculated. Determining candidate target points involves traversing all grid nodes within the mutation blind zone and selecting the node with the highest joint risk value as the representative re-sampling target for that blind zone. The estimated cost is calculated based on the aforementioned composite cost map, using a path planning algorithm (e.g., A* algorithm) to plan a path from the current position of the mobile reconnaissance platform to each candidate target point, and accumulating the travel cost of each grid along the path to obtain the estimated cost.

[0125] Specifically, the travel cost C(g) of each grid node g in the composite cost map is defined as: C(g) = Rg + Cbase, where Rg is the joint risk value of the node, and Cbase is the baseline travel cost (preset to 1.0, unit consistent with Rg). This design ensures that the travel cost in high-risk areas is greater than that in low-risk areas, and the path planning algorithm automatically tends to avoid risky areas. The planning of risky feasible paths uses the A* algorithm, with the cost function f(n) = g(n) + h(n), where g(n) is the actual cumulative travel cost from the starting point to the current node n along the planned path, g(n) = Σ{k∈path}C(k); h(n) is the heuristically estimated cost from the current node n to the target point, using an improved diagonal distance that considers the influence of wind field. Where dx and dy are the grid coordinate differences between node n and the target point. The wind field influence weight is preset to 0.15 and calibrated based on measured data of the wind field's impact on the aircraft's ground speed. This represents the wind speed value of the current wind field vector. It is the dot product of the direction vector from n to the target and the unit wind direction vector. It is set to 0 when the dot product is less than 0, that is, it only adds extra cost when flying against the wind.

[0126] The A* algorithm maintains an open list (a min-heap sorted by f(n) value) and a closed list. Each time, it takes the node with the smallest f(n) from the open list and expands it, extending its eight neighbors and updating the g and f values, until the target point is reached or the open list is empty. After the search is complete, it backtracks from the target point to its parent pointer to obtain the complete path. The estimated cost of the path is the g value (cumulative travel cost) of the target node. The input to this path planning is a composite cost map, the starting point coordinates (the current position of the mobile reconnaissance platform), the target point coordinates (candidate target points), and the current wind field vector. The output is an optimal path from the starting point to the target point and its estimated cost.

[0127] When there are candidate paths where the estimated consumption is less than or equal to the remaining endurance and the wind field vector is less than or equal to the safe operation threshold, the feasible blind zone with the highest replenishment priority index is selected as the mission target. This step aims to screen out feasible replenishment missions under current resource and environmental conditions.

[0128] The safe operation thresholds are preset based on the performance parameters of the mobile reconnaissance platform (such as maximum wind resistance and minimum flight altitude) and mission safety requirements, such as maximum permissible wind speed and maximum permissible gust intensity. The calculation of the supplementary mining priority index is based on a comprehensive assessment of factors such as the risk level, urgency, and area size of the abrupt change blind zone. For example, the joint risk value can be used as the priority index, or multiple indicators can be weighted and summed according to preset weights.

[0129] Otherwise, a multi-attribute evaluation matrix is ​​constructed based on the area, risk exposure value, and urgency of re-mining for all mutation blind zones. The re-mining priority index for each blind zone is calculated using the superior-inferior solution distance method, and feasible task objectives are reassigned from high to low based on the re-mining priority index. The risk exposure value is the sum of the updated joint risk values ​​of all grid nodes within the mutation blind zone. When no candidate path satisfies the above conditions, a more complex decision-making mechanism is employed. The multi-attribute evaluation matrix uses the area, risk exposure value, and re-mining urgency of each mutation blind zone as evaluation indicators.

[0130] The superior-inferior solution distance method, a multi-criteria decision analysis approach, evaluates the superiority or inferiority of each abruptly changed blind zone by calculating its distance from the ideal solution and the negative ideal solution, thereby obtaining a re-mining priority index. For example, different weights are assigned to area, risk exposure value, and re-mining urgency, and then the superior-inferior solution distance method is used to calculate the comprehensive priority score for each blind zone.

[0131] The mobile reconnaissance platform is controlled to perform supplementary data collection along the planned path and update the joint risk field in real time within a preset range from the target point. If a new risk intrudes into the planned path, a local replanning mechanism is triggered to avoid it. Once the mission objective is determined, the mobile reconnaissance platform will fly autonomously along the planned path. During flight, the platform continuously receives the latest monitoring data and updates its understanding of the joint risk field in real time. When the platform approaches the target point to a certain distance, or during flight, if a new high-risk area (i.e., a newly emerging risk) is detected that conflicts with or intrudes into the current planned path, a local replanning mechanism will be immediately triggered. For example, the mobile reconnaissance platform uses sensors such as infrared or visible light cameras to monitor the area ahead of the path in real time. Once new signs of fire or risk areas are detected, the path planning algorithm will be activated to adjust the flight path, avoid dangerous areas, and ensure mission safety and continuity.

[0132] If a new risk intrudes into the planned path, i.e., the shortest distance from the boundary of the new risk area to the current planned path is less than the platform's safe obstacle avoidance distance threshold, local replanning is triggered to avoid the obstacle. The safe obstacle avoidance distance threshold is determined as follows: Safe obstacle avoidance distance threshold = max(D_brake, D_sensor + D_margin), where D_brake = v² / (2·a_max) is the distance required for the platform to brake urgently at the current speed v and the maximum braking acceleration a_max; D_sensor is the effective detection range of the platform's obstacle avoidance sensors, which is the minimum value among all sensors, such as the effective detection range of the infrared camera being 200m; D_margin is the preset safety margin (initially set to 50m, which can be dynamically adjusted according to wind speed and visibility; for every 5m / s increase in wind speed, D_margin increases by 10m; for every 100m decrease in visibility, D_margin increases by 20m).

[0133] As can be seen from the above, this application first maps risk values ​​to access costs to generate a composite cost map, unifying the measurement standards for risk constraints and path planning. Then, combining the remaining range of the mobile platform and wind field safety constraints, it filters feasible replenishment paths and targets. When constraints are met, high-priority blind spots are directly selected for task execution. When constraints are not met, the replenishment order is scientifically ranked using a multi-attribute evaluation matrix and the superior-inferior solution distance method to maximize efficiency under resource constraints. The risk field is updated in real time during operation, and local replanning is triggered when new risks intrude into the path. This embodiment ensures both the scientific nature of replenishment decisions and resource utilization, while also possessing dynamic obstacle avoidance capabilities, improving the task efficiency and scheduling rationality of proactive replenishment while ensuring platform safety.

[0134] In one embodiment of this application, the urgency of supplementary sampling is calculated using the following logic: Extract the model cognitive bias of all grid nodes included in the current mutation blind zone after the most recent three consecutive time steps, calculate the time change rate of each node, and multiply the time change rate by the uncertainty measure of the current step of the same node to obtain the node-level dynamic risk entropy. Calculate the arithmetic sum of the dynamic risk entropies at all node levels within the mutation blind zone, and use it as the blind zone-level dynamic risk entropy; The shortest Euclidean distance between the current front-line boundary and the edge of the mutation blind zone is determined. Based on a preset negative exponential mapping function, the blind zone-level dynamic risk entropy and the shortest Euclidean distance are integrated into the urgency of re-mining.

[0135] In this embodiment, when calculating the urgency of re-collection, it is first necessary to extract the updated model cognitive bias of all grid nodes included in the current mutation blind zone over the most recent three consecutive time steps, and then calculate the rate of change over time for each node. Calculating the rate of change over time aims to capture the dynamic trend of this cognitive bias evolving over time. For example, it can be obtained by calculating the difference in updated model cognitive bias between adjacent time steps and then dividing by the time step size, i.e., (bias t - bias {t-1}) / Δt.

[0136] The calculation of node-level dynamic risk entropy aims to comprehensively consider the dynamics of model cognitive bias and the uncertainty of current predictions, thereby quantifying the risk level of a single grid node. For example, the calculated rate of change over time is multiplied by the updated uncertainty measure at the current time step.

[0137] Blind zone-level dynamic risk entropy is an aggregation of the risks of all nodes within the mutation blind zone, aiming to provide an overall quantitative indicator of risk for the entire mutation blind zone. This can be achieved by summing the node-level dynamic risk entropies of all grid nodes within the mutation blind zone.

[0138] Simultaneously, the shortest Euclidean distance between the current fire front boundary and the edge of the mutation blind zone is determined. The fire front boundary is the forefront of fire spread, while the edge of the mutation blind zone is the boundary of the identified high-risk area. This step aims to incorporate spatial geographic information to characterize the urgency of the fire threat to the blind zone. For example, using a distance transformation algorithm, the distance from each point within the mutation blind zone to the fire front boundary is first calculated, and then the minimum value among the points at the edge of the blind zone is taken.

[0139] The negative exponential mapping function is a nonlinear fusion function that converts distance into a decaying weight. This fusion aims to integrate dynamic risk and spatial distance to generate a final urgency level for replenishment, giving higher urgency to blind zones that are closer to the fire line and have higher dynamic risks. For example, the negative exponential mapping function is blind zone-level dynamic risk entropy × exp(-k × distance), where k controls the decay rate of urgency with respect to spatial distance. Therefore, the urgency level for replenishment is expressed as blind zone-level dynamic risk entropy × exp(-k × shortest Euclidean distance).

[0140] The attenuation coefficient k is determined as follows: Based on the average cruising speed v (m / s) and average data acquisition time tc (s) of the mobile reconnaissance platform, k = 2 / (v×tc) is set such that when the shortest Euclidean distance equals an effective operating radius (v×tc / 2), the urgency attenuates to its original value e. -1 ≈36.8%. For example, if the platform's cruising speed is 15 m / s and the average data collection time is 120 s, then the effective operating radius is 1800 m, k = 2 / 1800 ≈ 0.00111 m. -1 .

[0141] As can be seen from the above, this application first multiplies the time change rate of model cognitive bias with the uncertainty measure at each node to obtain the node-level dynamic risk entropy, while quantifying the deterioration speed and uncertainty of cognitive mismatch; then, it sums the dynamic risk entropy of all nodes in the blind zone to obtain the overall cognitive instability at the blind zone level; finally, it combines the shortest distance between the front of the fire and the blind zone, and integrates the imminent threat of the fire with the internal risk entropy through negative exponential mapping to generate the urgency of supplementary mining, further improving the rationality of supplementary mining scheduling and the accuracy of emergency response.

[0142] As can be seen from the above, when this application identifies a newly emerging risk intrusion planning path, it first estimates the risk diffusion envelope based on the fire behavior elliptical diffusion model and superimposes a safety buffer distance to form a dynamic no-fly zone, fully covering the potential range of fire spread. Then, starting from the platform's real-time position, it uses a dynamic A* algorithm to search for the optimal local path to avoid the no-fly zone on the composite cost map. If the original target point is covered by the risk, the suboptimal grid within the blind zone is automatically selected as a replacement target. This mechanism can quickly respond to sudden changes in the fire situation, dynamically adjust the flight path to avoid danger, and maximize the continuity of the target of the reconnaissance mission. Under the premise of ensuring the safety of the mobile reconnaissance platform operation, it improves the emergency adaptability and overall completion rate of mission execution.

[0143] Corresponding to the Bayesian variational inference-based air-space-ground forest fire monitoring method in the above embodiment, Figure 2 This is a structural block diagram of a space-air-ground forest fire monitoring system based on Bayesian variational inference, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The air-space-ground forest fire monitoring system 20 based on Bayesian variational inference includes: a data preprocessing module 21, a priori graph construction module 22, a Bayesian inference engine 23, a cognitive bias assessment module 24, a graph structure adaptation module 25, a risk field construction and identification module 26, an active supplementary sampling decision module 27, and a closed-loop iterative control module 28.

[0144] Among them, the data preprocessing module 21 is used to perform spatial registration and temporal alignment on the multi-source heterogeneous data of the target forest area, and generate multimodal feature vectors for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data and ground IoT sensor time series data; The prior graph construction module 22 is used to construct a dynamic spatiotemporal graph with grid nodes. Based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model, it generates the prior probability distribution of the fire status of each grid node and determines the graph structure of the dynamic spatiotemporal graph based on the prior probability distribution. The dynamic spatiotemporal graph contains grid nodes with multiple time steps, and the edges between nodes include spatial adjacency edges and temporal self-connection edges. The Bayesian inference engine 23 is used to input the dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node into a preset graph network based on factor graph and expectation propagation, and to calculate the posterior probability distribution representing the fire status of each grid node through Bayesian variational inference. The cognitive bias assessment module 24 is used to obtain the fire status estimate, uncertainty measure and model cognitive bias of each grid node based on the posterior probability distribution; the model cognitive bias is the difference between the expectation of the posterior probability distribution and the expectation of the prior probability distribution. The graph structure adaptive module 25 is used to use the node-by-node KL divergence between the posterior probability distribution and the prior probability distribution as the cognitive conflict degree, and to adjust the graph structure of the dynamic spatiotemporal graph according to the cognitive conflict degree. The Bayesian inference engine 23 is also used to re-execute Bayesian variational inference based on the dynamic spatiotemporal graph after adjusting the graph structure, to obtain an updated uncertainty measure and an updated model cognitive bias. The risk field construction and identification module 26 is used to construct a joint risk field based on the updated uncertainty metric and the updated model cognitive bias, and to identify mutation blind spots that meet the preset mutation conditions in the joint risk field. The active supplementary sampling decision module 27 is used to generate active supplementary sampling instructions based on the sudden blind zone and control the mobile reconnaissance platform to collect supplementary sampling data; The closed-loop iterative control module 28 is used to update the supplementary data to the grid nodes corresponding to the dynamic spatiotemporal graph, and trigger the graph structure adaptive module, Bayesian inference engine, risk field construction and identification module and active supplementary data acquisition decision module to repeatedly execute the complete closed-loop process from adjusting the graph structure of the dynamic spatiotemporal graph according to the degree of cognitive conflict to the identification of mutation blind zone, until the mutation blind zone is resolved.

[0145] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the data preprocessing module 21, prior graph construction module 22, Bayesian inference engine 23, cognitive bias assessment module 24, graph structure adaptive module 25, risk field construction and identification module 26, active supplementary sampling decision module 27, and closed-loop iterative control module 28 are shown.

[0146] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0147] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0148] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0149] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the air-space-ground forest fire monitoring method based on Bayesian variational inference provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0150] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0151] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0152] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0153] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0154] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0155] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0156] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0157] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring forest fires based on Bayesian variational inference, characterized in that, include: Spatial registration and temporal alignment are performed on the multi-source heterogeneous data of the target forest area to generate a multimodal feature vector for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data, and time-series data from ground IoT sensors; A dynamic spatiotemporal graph is constructed using the grid nodes. A prior probability distribution is generated based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model. The graph structure of the dynamic spatiotemporal graph is determined based on the prior probability distribution. The dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node are input into a preset graph network based on factor graph and expectation propagation, and the posterior probability distribution is obtained by Bayesian variational inference. Based on the posterior probability distribution, the fire status estimate, uncertainty measure, and model cognitive bias of the physical fire behavior model for each grid node are obtained. The cognitive conflict degree is determined by the node-wise KL divergence between the posterior probability distribution and the prior probability distribution, and the graph structure of the dynamic spatiotemporal graph is adjusted according to the cognitive conflict degree. Based on the dynamic spatiotemporal graph with the adjusted graph structure, Bayesian variational inference is re-executed to obtain updated uncertainty measures and updated model cognitive biases; A joint risk field is constructed based on the updated uncertainty metric and the updated model cognitive bias, and mutation blind spots that meet preset mutation conditions are identified in the joint risk field; Based on the aforementioned mutation blind zone, an active supplementary data collection command is generated to control the mobile reconnaissance platform to collect supplementary data. The supplementary data is updated to the grid nodes corresponding to the dynamic spatiotemporal map, and the complete closed-loop process, from adjusting the graph structure of the dynamic spatiotemporal map according to the degree of cognitive conflict to identifying mutation blind spots, is repeated until the mutation blind spots are resolved.

2. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 1, characterized in that, The process of generating a prior probability distribution of the fire status for each grid node based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model, and determining the graph structure of the dynamic spatiotemporal graph based on the prior probability distribution, includes: The digital elevation model, the meteorological forecast wind field, and the spatial distribution of combustible load are input into the physical fire behavior model for simulation. The input parameters are perturbed by Monte Carlo ensemble simulation to obtain the mean and variance of the fire intensity prediction for each grid node. The predicted mean and variance of the fire intensity are normalized, and the normalized mean and variance are used to construct a prior probability distribution of the fire status of the grid node characterized by a beta distribution. For each pair of spatially adjacent grid nodes, calculate the overlap coefficient between the two corresponding prior probability distributions; if and only if the overlap coefficient is greater than a preset overlap threshold, establish a directed spatial adjacency edge between the pair of spatially adjacent grid nodes, and set the edge weight of the directed spatial adjacency edge as the overlap coefficient. For each grid node, the Monte Carlo set is used to simulate the fire intensity prediction sample sequence generated at adjacent time steps, and the Pearson correlation coefficient of the fire intensity prediction sample sequence is calculated. The absolute value is taken as the time-series autocorrelation coefficient. A time-series self-connecting edge is established between the adjacent time steps, and the edge weight is set to the time-series autocorrelation coefficient.

3. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 1, characterized in that, The step of inputting the dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node into a preset graph network based on factor graphs and expectation propagation, and calculating the posterior probability distribution through Bayesian variational inference, includes: The prior probability distribution of each grid node is constructed as a prior factor, and the modal data in the multimodal feature vector is constructed as an observation factor based on a preset likelihood function. The edge weights carried by the spatial adjacency edges and temporal self-connection edges in the dynamic spatiotemporal graph are converted into smoothing factors between adjacent nodes. The prior factors, observed factors, and smoothing factors are respectively attached to the corresponding raster nodes and node pairs to generate a factor graph. The expectation propagation algorithm is used to perform message iteration on the factor graph: an approximate posterior distribution belonging to the beta distribution family is maintained for each node; In each iteration, for the prior factor and observed factor attached to a single node, the factor currently being processed is denoted as the target factor. The current message of the target factor is removed from the current approximate posterior distribution of the node to obtain the cavity distribution. The potential function carried by the target factor is combined to obtain the skew distribution. The skew distribution is projected onto the beta distribution family through moment matching to obtain the updated approximate posterior distribution. The message of the target factor passed to the variable node is then updated. When the changes in the approximate posterior distribution parameters of all nodes are less than the preset threshold, the iteration terminates, and the current approximate posterior distribution of each node is output as the posterior probability distribution.

4. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 3, characterized in that, The step of constructing observation factors from the modal data in the multimodal feature vector based on a preset likelihood function includes: For the same grid node, the likelihood of a fire is calculated using the Bernoulli likelihood model based on satellite hotspot monitoring data, UAV inspection video stream data, and ground IoT sensor time series data, respectively. Calculate the difference between each pair of the three sets of likelihood intensity values. If the difference between any two modal likelihood intensity values ​​is greater than the preset modal consistency threshold, obtain the false alarm rate and false negative rate of each modal data within the preset historical time window, and use the normalized false alarm rate and false negative rate to calculate the confidence weight of each modality. Based on the confidence weights, a weighted geometric mean is calculated on the likelihood values ​​of each modality to generate a joint likelihood intensity, and a joint observation factor is constructed based on the joint likelihood intensity. If the difference between the likelihood values ​​of any two modes is less than or equal to a preset modal consistency threshold, the arithmetic mean of the likelihood values ​​of each mode is directly calculated, and an observation factor is constructed based on the arithmetic mean.

5. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 1, characterized in that, The step of adjusting the graph structure of the dynamic spatiotemporal graph according to the degree of cognitive conflict includes: Grid nodes with cognitive conflict degree greater than a preset cognitive conflict threshold are marked as nodes to be reconstructed. All spatial adjacency edges connected to the nodes to be reconstructed in the dynamic spatiotemporal graph are removed. At the same time, the weight of the temporal self-connection edge corresponding to the node to be reconstructed is multiplied by the decay coefficient corresponding to the node to be reconstructed. For each node to be reconstructed, calculate the Helling distance between its posterior probability distribution and that of its spatially adjacent nodes. When the Helling distance is greater than the reconnection threshold, a directed spatial adjacency edge is established between the node to be reconstructed and the adjacent node, pointing from the node to be reconstructed to the adjacent node, and the edge weight is set to the Helling distance. For nodes whose cognitive conflict degree is less than or equal to the preset cognitive conflict threshold, their original spatial adjacency edges and temporal self-connection edges are retained, and the edge weights of the spatial adjacency edges and temporal self-connection edges are updated to the product of the current edge weight and the decay coefficient.

6. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 1, characterized in that, The preset mutation conditions include: The joint risk field is segmented by threshold, and a set of high-risk grids with joint risk values ​​greater than a preset risk threshold is extracted. Several high-risk connected components are also extracted based on spatial adjacency. For each high-risk connected component, for each grid node within that high-risk connected component, calculate the temporal change slope of the model cognitive bias over the most recent K consecutive time steps, and take the average of the absolute values ​​of the temporal change slopes of all grid nodes as the average change slope. Calculate the position difference vector of the geometric center within the high-risk connected domain between the current time step and K time steps ago, and take the absolute value of the projection of the position difference vector onto the unit vector of the prevailing wind direction as the cumulative displacement; If the average change slope is greater than a preset slope threshold, or the cumulative displacement is greater than a preset displacement threshold, then the high-risk connected component is determined to be a mutation blind zone.

7. The air-space-ground forest fire monitoring method based on Bayesian variational inference according to claim 1, characterized in that, The step of generating an active data collection instruction based on the mutation blind zone and controlling the mobile reconnaissance platform to collect supplementary data includes: The grids covered by the mutation blind zone in the joint risk field are mapped to access costs according to the joint risk value, and the areas outside the blind zone are also mapped to access costs according to the risk value. The composite cost map is then generated by merging these maps. Obtain the current location, remaining range, and current wind field vector of the mobile reconnaissance platform; For all identified mutation blind zones, the grid node with the largest joint risk value within each mutation blind zone is selected as a candidate target point; the estimated cost of the risky feasible path corresponding to each candidate target point is calculated. When there are candidate paths where the estimated consumption is less than or equal to the remaining endurance and the wind field vector is less than or equal to the safe operation threshold, the feasible blind area with the highest priority index for supplementary mining is selected as the task target. Otherwise, a multi-attribute evaluation matrix is ​​constructed based on the area, risk exposure value, and urgency of replenishment of all mutation blind zones. The replenishment priority index of each blind zone is calculated by the superior-inferior solution distance method, and feasible task objectives are reassigned from high to low based on the replenishment priority index. The risk exposure value is the sum of the updated joint risk values ​​of all grid nodes in the mutation blind zone. The mobile reconnaissance platform is controlled to perform supplementary data collection along the planned path and the joint risk field is updated in real time within a preset range from the target point. If a new risk intrudes into the planned path, local replanning is triggered to avoid it.

8. A space-air-ground forest fire monitoring system based on Bayesian variational inference, characterized in that, include: The data preprocessing module is used to perform spatial registration and temporal alignment on multi-source heterogeneous data of the target forest area to generate multimodal feature vectors for each grid node; the multi-source heterogeneous data includes satellite hotspot monitoring data, UAV inspection video stream data and time series data of ground IoT sensors; The prior graph construction module is used to construct a dynamic spatiotemporal graph with the grid nodes, generate a prior probability distribution of the fire status of each grid node based on the fire behavior state prediction and its uncertainty measure output by the physical fire behavior model, and determine the graph structure of the dynamic spatiotemporal graph based on the prior probability distribution. The Bayesian inference engine is used to input the dynamic spatiotemporal graph and the multimodal feature vectors corresponding to each grid node into a preset graph network based on factor graph and expectation propagation, and to calculate the posterior probability distribution representing the fire status of each grid node through Bayesian variational inference. The cognitive bias assessment module is used to obtain the fire status estimate, uncertainty measure, and model cognitive bias of each grid node based on the posterior probability distribution; the model cognitive bias is the difference between the expectation of the posterior probability distribution and the expectation of the prior probability distribution. The graph structure adaptive module is used to use the node-by-node KL divergence between the posterior probability distribution and the prior probability distribution as the cognitive conflict degree, and to adjust the graph structure of the dynamic spatiotemporal graph according to the cognitive conflict degree. The Bayesian inference engine is also used to re-execute Bayesian variational inference based on the dynamic spatiotemporal graph after adjusting the graph structure, to obtain an updated uncertainty measure and an updated model cognitive bias. The risk field construction and identification module is used to construct a joint risk field based on the updated uncertainty metric and the updated model cognitive bias, and to identify mutation blind zones that meet preset mutation conditions in the joint risk field. The active supplementary sampling decision module is used to generate active supplementary sampling instructions based on the mutation blind zone and control the mobile reconnaissance platform to collect supplementary sampling data; The closed-loop iterative control module is used to update the supplementary data to the grid nodes corresponding to the dynamic spatiotemporal graph, and trigger the graph structure adaptive module, Bayesian inference engine, risk field construction and identification module and active supplementary data acquisition decision module to repeatedly execute the complete closed-loop process from adjusting the graph structure of the dynamic spatiotemporal graph according to the degree of cognitive conflict to the identification of mutation blind zone, until the mutation blind zone is resolved.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images